How a call runs

#From step to step

1. A call starts at start_node.
2. When a step finishes, every step in its next starts. next can be one name or a list.
3. A Branch picks its own next step from mappings instead. If no path matches, it falls back to a default path, then to its next.
4. A step with no next (and no matching path) ends that line of work. When nothing is left running, the agent waits for the caller.

#Steps side by side

When next lists several steps, they all start at once. This is how you show the caller's words in the transcript while the AI thinks, or play a quick acknowledgement while a slower model works:

1"ask_for_input": {
2  "type": "input",
3  "parameters": { "input_variables": { "user_input": "str" } },
4  "next": ["ack", "main_reply", "transcription"]
5}

When two or more AI reply steps run side by side, the one that starts speaking first is heard first. The others are held, then played in the order they started. Give a held step a stream_timeout(seconds) and it's dropped if it's still waiting after that long, so a stale answer is never spoken.

#Waiting for the caller

A Listen step pauses its line of the workflow until the caller has finished speaking. The words are saved (normally as user_input) and the workflow carries on from the Listen step's next. Every turn of a conversation is one pass from Listen round to Listen again.

#When a step fails

If an AI provider fails or times out, the engine retries once on the next provider (see Models and failover). If that fails too, the agent says its fallback line. Any other failing step is skipped and logged; the rest of the workflow keeps going.

#Next steps

• Values and state: what steps read and write.
• Patterns: shapes that work well.